The AI era has produced a clear financial ranking, and chipmakers now sit at the top. NVIDIA’s Q2 fiscal 2027 results, posted August 26, confirmed what 2026’s equity returns had already revealed. The company reported $96.2 billion in quarterly revenue, the largest single quarter in semiconductor history, with $89 billion from data center (+117% year over year), 75% gross margins, and $59.7 billion in net income (+126% YoY). The $108 billion Q3 guide suggests the divergence between semiconductor producers and Big Tech buyers is not narrowing. It is widening.
Why chipmakers are winning the AI era
The mechanics behind the rotation are structural. The Magnificent Seven, Alphabet, Amazon, Apple, Microsoft, Meta, NVIDIA, and Tesla, drove S&P 500 returns in 2023 and 2024 through platform dominance and cloud scale. In 2026, those same companies are the largest buyers of AI compute. Combined AI-related capex by hyperscale cloud providers is projected to surpass $730 billion in 2026, and every dollar of that spending flows directly to chipmakers building the GPUs, HBM, and networking silicon.
The result is a transfer of return on capital from platform companies funding AI infrastructure to the hardware companies enabling it. Chip stocks reached $9.57 trillion in combined market capitalization in July 2026, accounting for 13.9% of the S&P 500, up from a much smaller share just two years earlier. The more Microsoft, Meta, and Amazon spend on data centers, the more revenue accrues to NVIDIA, Micron, SK Hynix, Broadcom, and Marvell. The pattern is not a trade; it is the AI boom’s underlying economics.
The semiconductor stocks vs Magnificent Seven divergence
The numbers leave little room for interpretation. Micron Technology is up 220% year to date. Marvell Technology is up 185%. Intel is up 150%. A popular semiconductor ETF has gained more than 70% in 2026. The Magnificent Seven ETF is up just 4%. Meta shares have declined from recent peaks despite billions poured into AI infrastructure, and Microsoft has not set a record high in more than ten months. Alphabet and Amazon are up roughly 8% and 11% respectively, both retracing from earlier peaks.
Matt Maley, chief market strategist at Miller Tabak + Co., acknowledged the durability of the dynamic while flagging risk. “We have seen other cracks over the past year, and they have not upset the apple cart for very long,” he said. “So it would be foolish to say the AI bubble is about to burst.” Investors are watching how those cracks develop, but the supply-constrained nature of the chipmakers’ pricing power continues to insulate the sector.
How NVIDIA extracts more revenue per gigawatt of demand
NVIDIA sits at the center of this rotation for technical reasons that go beyond market share. Each generation of its AI architecture expands the dollar value extracted from a given amount of computing capacity. The Hopper-era platform generated roughly $18 billion of revenue per gigawatt of deployed data center capacity. Blackwell expanded that to approximately $25 billion per gigawatt. The Vera Rubin platform, now in full production, is estimated at approximately $40 billion per gigawatt because it bundles the Vera CPU, Rubin GPU, NVLink 6 switching, ConnectX-9 networking, and software into a single integrated purchase.
The Vera CPU that NVIDIA VP Ian Buck hand-delivered to AWS on August 27 is the newest element of that bundle. It carries 88 custom Olympus cores, 1.2 terabytes per second of memory bandwidth, and a threading architecture called Spatial Multithreading. CFO Colette Kress told investors the Vera CPU opens roughly $200 billion in addressable CPU revenue. Supply commitments grew from $119 billion at the end of Q1 to $279 billion at the end of Q2, primarily driven by procurement of high-bandwidth memory for Vera Rubin.
The HBM bottleneck powering chipmakers’ gains
High-bandwidth memory is the specific technology at the center of the constraint. Standard server memory delivers data at roughly 50 to 100 gigabytes per second per module. HBM4, used in NVIDIA’s Vera Rubin platform, delivers more than 1.4 terabytes per second per stack by stacking memory dies vertically and bonding them directly alongside the GPU on a silicon interposer. The HBM bandwidth advantage, roughly 15 to 20 times that of conventional DRAM, is what makes large-scale AI model inference commercially viable.
The supply constraint is structural. Producing one gigabyte of HBM consumes approximately three to four times the semiconductor wafer area of a gigabyte of standard DDR5, and every HBM stack must be assembled using TSMC’s CoWoS advanced packaging. NVIDIA holds roughly 60% of available CoWoS packaging capacity for its data center accelerators in 2026. SK Hynix’s CEO has stated publicly that 2027 will be the worst year of the HBM shortage, with demand expected to outpace supply beyond 2030. Gartner’s August 24, 2026 forecast projected DRAM revenue to rise 246.6% in 2026, with total memory revenue surpassing $1 trillion by 2027.
Wall Street finances AI like a toll road
The most consequential structural development of NVIDIA’s earnings week was not the revenue figure. On August 10, NVIDIA signed memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to establish AI compute financing platforms designed to mobilize more than $500 billion of third-party institutional capital for AI infrastructure. The structure treats AI compute as an investable infrastructure asset, similar to how Wall Street finances toll roads, cell towers, and power plants.
Apollo President Jim Zelter called AI compute a “scarce, mission-critical asset class.” BlackRock CEO Larry Fink framed compute infrastructure as a long-duration economic asset. The deal’s timing was explicit: a July 2026 market pullback raised questions about whether Big Tech’s AI spending would generate commensurate returns, and the $500 billion financing structure was NVIDIA’s answer. The risk is that AI GPUs depreciate in three to five years, far faster than the 25-to-40-year lifespan of traditional infrastructure assets, a gap that will test whether institutional capital accepts compute as durable collateral.
The China zero and the durability of the rotation
Embedded in NVIDIA’s $108 billion Q3 guidance is a disclosure that often gets less attention than the headline. That forecast assumed essentially zero data center compute revenue from China. In Q2, China data center revenue represented less than 1% of segment revenue. The US export restrictions have effectively severed GPU supply to Chinese data centers, and the January 2026 H200 clearances for Alibaba, ByteDance, and JD.com have not translated into significant shipments. NVIDIA is, however, accepting orders for its Vera CPU from Chinese customers, with deliveries targeted for August 2026.
The structural answer on durability is at least through 2027. SK Hynix’s forecast that HBM demand will outpace supply through the end of the decade and NVIDIA’s own supply-constrained 70% guidance when demand implies 140% both confirm that scarcity is operational, not speculative. Every major AI infrastructure expansion planned for 2027 and 2028 is being priced against a supply ceiling set by HBM production, not customer willingness to pay. The hand-delivery of NVIDIA’s first Vera CPU to AWS in Seattle is both symbol and data point: chipmakers are physically bringing the next architecture to the cloud’s most important customers, one system at a time, at the moment when every major financial institution has decided compute infrastructure is the decade’s most important investable asset class. chipmakers